Process planning model pre-training method and system based on multi-processing scene segmentation imitation learning

Abstract

This invention discloses a pre-training method and system for process planning models based on segmented imitation learning across multiple processing scenarios, relating to the field of computer-aided manufacturing. The method involves generating processing paths with various strategies using a mature process database of CAM software for multiple processing scenarios with different geometric features, material properties, and processing requirements. A large number of initial states are generated through virtual trial cutting using a processing world model, providing rich and diverse training samples for reinforcement learning pre-training. Based on this, a segmented behavior cloning method is used for large-scale pre-training, resulting in a process autonomous planning base agent that has learned processing strategies under different processing scenarios. When faced with new processing tasks, processing paths can be quickly generated through transfer learning. This invention significantly improves generalization ability while achieving geometric-physical-control collaborative planning, providing an intelligent solution for efficient process planning in modern manufacturing.

Description

Technical Field

[0001] This invention relates to the field of computer-aided manufacturing, and in particular to a method and system for pre-training process planning models based on segmented imitation learning across multiple processing scenarios. Background Technology

[0002] In the field of computer-aided manufacturing (CAM), traditional process planning techniques have long relied on path generation methods based on geometric features. Their core limitation lies in focusing solely on the workpiece's geometry, neglecting the combined influence of physical factors such as material properties, cutting dynamics, machine tool dynamic response, and CNC control parameters. This single-dimensional planning approach necessitates extensive trial cutting verification and manual adjustments in practical applications, resulting in high costs and poor process consistency for CAM-based process planning and optimization.

[0003] Reinforcement learning-based autonomous process planning agents, as an intelligent planning method, autonomously learn optimal process strategies through interaction with the environment, achieving collaborative planning of geometry, physics, and control. However, their practical application in manufacturing environments faces the dual challenges of low training efficiency and weak generalization ability. Reinforcement learning agents require extensive interaction with the real manufacturing environment to collect training data, a process that is both time-consuming and expensive in actual production environments. Furthermore, models trained for a single processing scenario struggle to adapt to changes in other scenarios. When geometric features, material properties, or processing requirements change, large-scale retraining is often necessary, leading to high application costs and severely limiting their practical value in complex manufacturing environments.

[0004] To address these issues, there is an urgent need for a pre-training method and system for process planning models based on segmented imitation learning across multiple processing scenarios. Summary of the Invention

[0005] To address the aforementioned issues, this application proposes a pre-training method for process planning models based on segmented imitation learning across multiple processing scenarios, the details of which are as follows:

[0006] S1. Obtain the processing scene and use CAM software to perform process planning to obtain the processing path dataset;

[0007] S2. Based on the processing progress information and key feature information, segment the processing path data in the processing path dataset and input it into the processing process world model to perform virtual trial cutting to obtain the initial state data.

[0008] S3. Based on the imitation learning method of reinforcement learning, the initial state data is trained and the collaborative planning capabilities of geometry, physics and control are integrated to obtain the process planning strategy.

[0009] Preferably, the specific content of the processing path dataset obtained in S1 by acquiring the processing scene and using CAM software to perform process planning on the processing scene includes:

[0010] The processing scenarios include planar processing scenarios, cavity processing scenarios, freeform surface processing scenarios, and blade processing scenarios;

[0011] The processing scenarios constitute a set of processing scenarios. ;

[0012] CAM software integrates key machining features and standard milling process methods during the process planning of machining scenarios;

[0013] For processing scenarios The complete machining path is generated by CAM. , i represents the set of processing scenarios The processing scene number in the middle, Representing a scene The length of the path sequence generated by CAM.

[0014] Preferably, in S2, the processing path data in the processing path dataset is segmented based on processing progress information and key feature information, and then input into the processing process world model for virtual trial cutting to obtain the initial state data. The specific content of this process is as follows:

[0015] Preset breakpoints are generated based on processing progress information and key feature information. ,in, This represents the number of segments in this processing scenario;

[0016] According to the preset breakpoint An initial state data is obtained by segmenting a processing path data in the processing path dataset. ;

[0017] Several initial state data constitute the initial state library ,in, l Represents the total number of initial states. .

[0018] Preferably, the imitation learning method based on reinforcement learning in S3, which trains on the initial state data and integrates the collaborative planning capabilities of geometry, physics, and control to obtain the process planning strategy, includes the following specific content:

[0019] The reinforcement learning-based imitation learning method introduces the process of reinforcement learning and environment interaction learning into imitation learning;

[0020] The training process using initial state data employs a closed-loop iterative mechanism and introduces a dynamic enhancement sampling strategy, utilizing quantitative indicators such as surface quality and cutting force stability. Evaluate the strategy performance in each initial processing scenario And update the sampling probability distribution ,in, i The largest category is based on the corresponding processing scenario. j This corresponds to the number of a specific segment in this processing scenario. It is used to evaluate the overall performance of the segment and to sample the probability of scenarios that do not achieve the target effect. Adjustments have been made, including changes to the subscript logic as described above. Similarly, the 0 in the subscript here represents the initial state, which is a commonly used marker in reinforcement learning and imitation learning.

[0021] Preferably, the specific process of training on the initial state data includes:

[0022] S301. Sample the initial state, i.e., from the initial state library. A mid-sample yields an initial state. initialize the environment to ;

[0023] S302, Experience Accumulation Moment, for state Utilize the current strategy Sampling to obtain action The action is then passed into the environment for execution, yielding the next state and corresponding reward. ,in The state transition function of the environment;

[0024] S303, Current Status Current action ,award And the next step status This is called an experience and is stored in the experience revisit buffer. middle;

[0025] S304, Calculation Strategy initial state Overall performance and according to relative to other initial environments It's worth updating sampling probability ;

[0026] S305. Repeat S301-S304 to obtain a large amount of data for the imitation learning training of the process planning model.

[0027] S306. A strategy is implemented using an offline reinforcement learning method based on SAC. The update ultimately yields the action value function. and strategy Output.

[0028] Preferably, the specific content of S306 is as follows:

[0029] S3061, Update Action Value Function ,Right now Then, update the strategy based on the gradient of the action value function. ,Right now ;

[0030] S3062, Update the temperature term used to regulate the strategy exploration rate during experience accumulation, i.e. ;

[0031] S3063, Last Update Target Action Value Function Its function is to lag behind Updates are made to ensure the stability of the action value function; the update method is as follows. .

[0032] S3064, repeat S3061-S3063, until the process planning model is established. Convergence, the criteria for determining convergence require a manual observation strategy. The loss curve continues to decrease until it stops decreasing, ultimately yielding the action value function. and strategy Output.

[0033] Preferably, a PPO-based transfer learning fine-tuning mechanism is introduced for training on the initial state data, specifically including:

[0034] In the transfer learning process, the initial state of the target processing scenario ,by As an initial strategy As an initial value function, it interacts online with the target processing environment;

[0035] The agent generates a processing path based on the current policy. The environment provides immediate rewards, and the policy is optimized using a substitute objective function of PPO.

[0036] A pre-training system for process planning models based on segmented imitation learning across multiple processing scenarios includes:

[0037] Processing scene selection unit: Obtain the processing scene, and use CAM software to plan the process of the processing scene to obtain the processing path dataset;

[0038] Initial state construction unit: Based on the processing progress information and key feature information, the processing path data in the processing path dataset is segmented and input into the processing process world model to obtain the initial state data through virtual trial cutting;

[0039] Training strategy generation unit: Based on the imitation learning method of reinforcement learning, it trains on the initial state data and integrates the collaborative planning capabilities of geometry, physics and control to obtain the process planning strategy.

[0040] In summary, the process planning model pre-training method and system based on segmented imitation learning in multiple processing scenarios of this invention, compared with traditional technologies, targets multiple processing scenarios with different geometric features, material properties, and processing requirements. It utilizes the mature process database of CAM software to generate processing paths with various strategies, and generates a large number of initial states through virtual trial cutting using a processing world model, providing rich and diverse training samples for reinforcement learning pre-training. Based on this, a segmented behavior cloning method is used for large-scale pre-training, resulting in a process autonomous planning base agent that has learned processing strategies under different processing scenarios. When faced with new processing tasks, it can quickly generate processing paths through transfer learning. This method combines the advantages of reinforcement learning and CAM, significantly improving generalization ability while achieving geometric-physical-control collaborative planning, providing an intelligent solution for efficient process planning in modern manufacturing.

[0041] The technical method of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0042] Figure 1 This is a framework diagram of the process planning model pre-training method based on segmented imitation learning in multiple processing scenarios according to the present invention;

[0043] Figure 2 This is a schematic diagram of the processing scene in an embodiment of the present invention;

[0044] Figure 3 This is a simulation diagram of material removal in an embodiment of the present invention;

[0045] Figure 4 This is a schematic diagram of cutting force simulation in an embodiment of the present invention. Figure 4 (a) in the figure represents the simulation of the cutting force in the X direction. Figure 4 (b) in the figure represents the simulation of the cutting force in the Y direction. Figure 4 (c) in the figure represents the simulation of the cutting force in the Z direction;

[0046] Figure 5 This is a schematic diagram of bending moment wear simulation in an embodiment of the present invention. Figure 5 (a) in the figure represents the bending moment simulation. Figure 5 (b) in the figure represents the wear simulation. Detailed Implementation

[0047] The technical method of the present invention will be further described below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of this application.

[0048] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the scope of this application and its application or use.

[0049] Techniques, systems, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, they should be considered part of the instruction manual.

[0050] In all the examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0051] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0052] Example 1

[0053] This invention provides a pre-training method for process planning models based on segmented imitation learning across multiple processing scenarios, such as... Figure 1 As shown, a process planning intelligent agent base model with strong generalization ability will be constructed through a systematic process approach. This process planning intelligent agent base model can effectively cope with new processing tasks and can quickly generate high-quality processing paths through simple transfer learning, significantly reducing the development cost and time of process planning and providing efficient and intelligent process planning solutions for the manufacturing industry.

[0054] The details are as follows:

[0055] S1. Obtain the machining scene. Use CAM software to perform process planning on the machining scene to obtain the machining path dataset. These datasets not only contain geometric path information, but also integrate cutting parameters and process conditions in the actual machining process, providing high-quality basic samples for subsequent training.

[0056] Furthermore, in S1, the processing scene is acquired, and CAM software is used to perform process planning on the processing scene. The specific content of the diverse processing path dataset covering different geometric features, material properties, and processing requirements includes:

[0057] The processing scenarios include typical processing scenarios such as planar processing scenarios, cavity processing scenarios, freeform surface processing scenarios, and blade processing scenarios.

[0058] The processing scenarios constitute a set of processing scenarios. .

[0059] CAM software integrates key machining features and standard milling process methods during the process planning of machining scenarios, such as... Figure 2 As shown, the selected workpieces need to have key machining features such as cavity milling, end milling, side milling, and grooving evenly distributed to avoid training bias caused by feature concentration. At the same time, it integrates standard milling processes such as helical milling and zigzag milling, providing high diversity and high representativeness for the process data generated by CAM software, laying the foundation for the intelligent agent to systematically learn the process rules of multiple machining scenarios.

[0060] Figure 3 Figure 4 and Figure 5 Corresponding to Figure 2 The simulation includes material removal, cutting force, bending moment, and wear simulations, all of which are performed simultaneously.

[0061] For processing scenarios The complete machining path is generated by CAM. , i represents the set of processing scenarios The processing scene number in the middle, Representing a scene The length of the path sequence generated by CAM.

[0062] A complete processing path typically contains tens of thousands or even hundreds of thousands of control points, covering different processing features. Directly using the entire path as a training set for imitation learning makes it difficult for the learning algorithm to effectively extract key features and establish effective decision mappings, resulting in difficulty in algorithm convergence and poor training performance.

[0063] To address this issue, this invention employs a method of segmenting the complete processing path according to processing progress and processing characteristics and constructing corresponding initial training environments to improve the training effect of imitation learning and the model's generalization ability.

[0064] Specifically, S2, based on the processing progress information and key feature information, the processing path data in the processing path dataset is reasonably segmented and input into the processing process world model to perform virtual trial cutting, thereby obtaining a large amount of initial state data containing various typical features and different processing progress for imitation learning.

[0065] In S2, the processing path data in the processing path dataset is segmented based on processing progress information and key feature information, and then input into the processing process world model for virtual trial cutting to obtain the initial state data. The specific content of this data is as follows:

[0066] For processing scenarios The complete machining path is generated by CAM. , i represents the set of processing scenarios The processing scene number in the middle, Representing a scene The path sequence length generated by CAM is input into the world model of the machining process for virtual trial cutting.

[0067] During the virtual trial cutting process, preset breakpoints are generated based on processing progress information and key feature information. ,in, This represents the number of segments in this processing scenario.

[0068] According to the preset breakpoint An initial state data is obtained by segmenting a processing path data in the processing path dataset. .

[0069] Several initial state data constitute the initial state library ,in, l Represents the total number of initial states. .

[0070] For example, a breakpoint is defined as each certain number of steps taken or each specific feature processed. Each time a breakpoint is reached... The system stores the current geometric state of the workpiece. (Including the precise shape and dimensions of the machined portion) and the relative position of the cutting tools. Cutting parameters Key information such as feed rate and spindle speed is used to form a complete initial state. These initial states not only contain processing progress information, but also integrate physical characteristics and control parameters, providing high-quality training samples for imitation learning.

[0071] In this way, a complete processing path It is transformed into several independent initial states with definite physical meaning. This allows the agent to focus on learning decision-making strategies for specific processing stages, rather than performing global optimization of the entire complex path all at once. This segmented approach effectively reduces the complexity of the state space, enabling the agent to more accurately capture key features of different processing stages, thereby significantly improving the efficiency and effectiveness of imitation learning across all processing paths. By performing the same operation, an initial state library containing various processing scenarios and features can be obtained. .

[0072] In the imitation learning training phase, this invention proposes a "reinforcement learning-based imitation learning training" method. By introducing the process of reinforcement learning and environment interaction learning into imitation learning, it effectively solves the data volume bottleneck and transferability problems of traditional imitation learning methods. Traditional imitation learning methods rely on a large amount of high-quality training data. Although CAM software can generate process path data, its quantity is insufficient to meet the needs of large-scale training. Furthermore, due to the high complexity and precision of the processing scenario, conventional data augmentation methods (such as path perturbation or parameter transformation) are prone to introducing erroneous data, causing the model to learn unreasonable processing strategies and affecting the final performance.

[0073] More importantly, traditional imitation learning only learns the mapping from state to action, lacking the support of a value network. This makes it impossible for the trained model to undergo subsequent reinforcement learning optimization and transfer, severely restricting its application ability in new scenarios. This invention enables the agent to explore autonomously in a virtual environment through a reinforcement learning mechanism, dynamically generating diverse training samples, and simultaneously constructing a value network to provide a theoretical basis for policy optimization, thereby achieving a seamless transition from imitation learning to reinforcement learning.

[0074] S3. The imitation learning method based on reinforcement learning, through the interaction between the agent and the environment with different initial states, trains and learns the initial state data under various processing features and processing progress, and integrates the collaborative planning ability of geometry-physics-control to enable the agent to master the process rules under different scenarios during the pre-training process, forming a process planning strategy with strong generalization ability.

[0075] Furthermore, the imitation learning method based on reinforcement learning in S3, which trains on initial state data and integrates geometric-physical-control collaborative planning capabilities to obtain the specific content of the process planning strategy, includes:

[0076] The reinforcement learning-based imitation learning method introduces the process of reinforcement learning and environment interaction learning into imitation learning;

[0077] The training process using initial state data employs a closed-loop iterative mechanism, with the system starting from a pre-built initial state library. A scene is randomly loaded in the middle. intelligent agent Interact with the scene to gain experience Each contains a state. ,action ,award and the state at the next moment These experiences are stored in the experience replay buffer. In each training round, the system samples a batch of empirical data from the buffer and first updates the value network by minimizing the Bellman error. This enables it to accurately predict long-term cumulative returns under different states; subsequently, the updated value network is used to calculate the policy gradient. For strategy models Optimization is carried out to gradually improve the decision-making quality of the agent in the geometric-physical-control collaborative dimension. This mechanism ensures that the agent not only learns expert strategies, but also masters the value assessment ability at different processing stages, laying the foundation for subsequent transfer training.

[0078] To improve training efficiency and data utilization value, a dynamic augmentation sampling strategy was introduced into the training process using initial state data, employing quantitative indicators such as surface quality and cutting force stability. Evaluate the strategy performance in each initial processing scenario And update the sampling probability distribution Increase the sampling probability for scenarios with poor performance. This targeted data augmentation mechanism enables the model to optimize inefficient regions more effectively, improve data utilization, and accelerate the convergence process. In this way, the effective utilization of training data is significantly improved, and the model can master a comprehensive strategy covering multiple geometric features, material properties, and processing requirements in a short period of time, providing a reliable foundation for rapid migration when facing new processing tasks.

[0079] Furthermore, the algorithm requires several iterations. For each iteration, the specific process of training on the initial state data includes:

[0080] S301. Sample the initial state, i.e., from the initial state library. A mid-sample yields an initial state. initialize the environment to .

[0081] S302. Experience accumulation, regarding the state Utilize the current strategy Sampling to obtain action The action is then passed into the environment for execution, yielding the next state and corresponding reward. ,in This is the state transition function for the environment.

[0082] S303, Current Status Current action ,award And the next step status This is called an experience and is stored in the experience revisit buffer. In the middle; by repeatedly executing such a "sampling-execution-store" loop, a large amount of experience can be obtained. The process of acquiring experience is synchronized with S305 until the strategy in S305 is implemented. Convergence can stop the acquisition of experience.

[0083] After accumulating experience with the initial state, dynamic enhancement sampling will be evaluated to improve the sampling probability of the poorly performing initial environment, thereby enhancing the effectiveness of experience accumulation.

[0084] S304, Calculation Strategy initial state Overall performance and according to relative to other initial environments It's worth updating sampling probability .

[0085] S305. Repeat S301-S304 to obtain a large amount of data for the imitation learning training of the process planning model.

[0086] S306. A strategy is implemented using an offline reinforcement learning method based on SAC. The update ultimately yields the action value function. and strategy Output.

[0087] Furthermore, the specific details of S306 are as follows:

[0088] S3061, Update Action Value Function ,Right now Then, update the strategy based on the gradient of the action value function. ,Right now .

[0089] S3062, Update the temperature term used to regulate the strategy exploration rate during experience accumulation, i.e. .

[0090] S3063, Last Update Target Action Value Function Its function is to lag behind Updates are made to ensure the stability of the action value function; the update method is as follows. .

[0091] S3064, repeat S3061-S3063, until the process planning model is established. Convergence, the criteria for determining convergence require a manual observation strategy. The loss curve continues to decrease until it stops decreasing, ultimately yielding the action value function. and strategy The output and pseudocode of the algorithm are shown below:

[0092]

[0093] In practical processing applications, although the process planning base model constructed by imitation learning has a wide generalization ability, its processing performance is still insufficient when facing specific workpieces, materials or precision requirements. To solve this problem, this invention introduces a PPO-based transfer learning fine-tuning mechanism for training initial state data. The base model is fine-tuned for specific processing tasks to obtain better adaptability and performance. The PPO algorithm is an ideal choice for fine-tuning due to its excellent stability. By introducing a policy update pruning mechanism, it effectively prevents the training instability problem caused by excessive policy update amplitude, while maintaining high sample efficiency.

[0094] The specific content includes:

[0095] In the transfer learning process, the initial state of the target processing scenario ,by As an initial strategy As an initial value function, it interacts online with the target processing environment.

[0096] The agent generates a processing path based on the current policy. The environment provides immediate rewards, and the policy is optimized using a substitute objective function of PPO.

[0097] This PPO-based fine-tuning process requires no large amount of new data and only minimal environmental interaction to achieve efficient optimization, significantly improving the targeting and accuracy of process planning and ensuring that the model achieves optimal performance in practical applications. The specific algorithm is shown below:

[0098]

[0099] A pre-training system for process planning models based on segmented imitation learning across multiple processing scenarios includes:

[0100] Processing scene selection unit: Obtain the processing scene, and use CAM software to perform process planning on the processing scene to obtain the processing path dataset.

[0101] Initial state construction unit: Based on the processing progress information and key feature information, the processing path data in the processing path dataset is segmented and input into the processing process world model to obtain the initial state data through virtual trial cutting.

[0102] Training strategy generation unit: Based on the imitation learning method of reinforcement learning, it trains on the initial state data and integrates the collaborative planning capabilities of geometry, physics and control to obtain the process planning strategy.

[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical methods of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical methods of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical methods to deviate from the spirit and scope of the technical methods of the present invention.

Claims

1. A process planning model pre-training method based on multi-processing scene segmentation imitation learning, characterized in that, Includes the following steps: S1. Obtain the processing scene and use CAM software to perform process planning to obtain the processing path dataset; S2. Based on the processing progress information and key feature information, segment the processing path data in the processing path dataset and input it into the processing process world model to perform virtual trial cutting to obtain the initial state data. S3. Based on reinforcement learning, the imitation learning method trains the initial state data and integrates the collaborative planning capabilities of geometry, physics and control to obtain the process planning strategy. The specific process of training on the initial state data includes: S301. Sample the initial state, i.e., from the initial state library. A mid-sample yields an initial state. initialize the environment to , This represents the sampling probability distribution; S302, Experience Accumulation Moment, for state Utilize the current strategy Sampling to obtain action The action is then passed into the environment for execution, yielding the next state and corresponding reward. ,in The state transition function of the environment; S303, Current Status Current action ,award And the next step status This is called an experience and is stored in the experience revisit buffer. middle; S304, Calculation Strategy initial state Overall performance and according to relative to other initial environments It's worth updating sampling probability ; S305. Repeat S301-S304 to obtain a large amount of data for the imitation learning training of the process planning model. S306. A strategy is implemented using an offline reinforcement learning method based on SAC. The update yields the action value function. and strategy Output; The specific content of S306 is as follows: S3061, Update Action Value Function ,Right now Then, update the strategy based on the gradient of the action value function. ,Right now ; S3062、update the temperature term used to regulate the policy exploration degree when accumulating experience, that is ; S3063, last update target action value function , which functions to lag behind updating to ensure the stability of the action value function, and the updating method is ; S3064, repeat S3061-S3063, until the process planning model is established. Convergence yields the action value function. and strategy Output; A PPO-based transfer learning fine-tuning mechanism is introduced for training on initial state data. The specific content includes: In the transfer learning process, the initial state of the target processing scenario ,by As an initial strategy As an initial value function, it interacts online with the target processing environment; The agent generates a machining path according to a current strategy , environment feedback instant reward, and strategy optimization through the alternative objective function of PPO.

2. The process planning model pre-training method based on multi-workshop scenario segmentation imitation learning according to claim 1, characterized in that, The specific content of the processing path dataset obtained in S1, which is derived from the processing scene using CAM software for process planning, includes: The processing scenarios include planar processing scenarios, cavity processing scenarios, freeform surface processing scenarios, and blade processing scenarios; The processing scene constitutes a processing scene set ; CAM software integrates key machining features and standard milling process methods during the process planning of machining scenarios; For processing scenarios The complete machining path is generated by CAM. i represents the set of processing scenarios The processing scene number in the middle, Representing a scene The length of the path sequence generated by CAM.

3. The process planning model pre-training method based on multi-workshop scenario segmentation imitation learning according to claim 2, characterized in that, In S2, the processing path data in the processing path dataset is segmented based on processing progress information and key feature information, and then input into the processing process world model for virtual trial cutting to obtain the initial state data. The specific content of this data is as follows: Preset breakpoints are generated based on processing progress information and key feature information. ,in, This represents the number of segments in this processing scenario; According to a preset breakpoint Segmenting one piece of machining path data in the machining path data set to obtain an initial state data ; A number of initial state data constitute an initial state library wherein, l representing the total number of initial states, .

4. The process planning model pre-training method based on multi-workshop scenario segmentation imitation learning according to claim 3, characterized in that, The imitation learning method based on reinforcement learning in S3, which trains on initial state data and integrates geometric-physical-control collaborative planning capabilities to obtain the process planning strategy, includes the following specific content: The reinforcement learning-based imitation learning method introduces the process of reinforcement learning and environment interaction learning into imitation learning; The training process using initial state data employs a closed-loop iterative mechanism and introduces a dynamic enhancement sampling strategy, using surface quality and cutting force stability as quantitative indicators. Evaluate the strategy performance in each initial processing scenario And update the sampling probability distribution Sampling probability for scenarios that do not achieve the target effect Adjustments will be made.

5. A process planning model pre-training system based on multi-processing scenario segmented imitation learning, applied to the process planning model pre-training method based on multi-processing scenario segmented imitation learning as described in any one of claims 1-4, characterized in that, include: Processing scene selection unit: Obtain the processing scene, and use CAM software to plan the process of the processing scene to obtain the processing path dataset; Initial state construction unit: Based on the processing progress information and key feature information, the processing path data in the processing path dataset is segmented and input into the processing process world model to obtain the initial state data through virtual trial cutting; Training strategy generation unit: Based on the imitation learning method of reinforcement learning, it trains the initial state data and integrates the collaborative planning capabilities of geometry, physics and control to obtain the process planning strategy.

Images